Anthropic's Claude Code team has published a detailed breakdown of how it reviews code when an AI writes roughly 90 percent of it. The answer is not "humans stopped reading" — it's that reading moved from a single human gate at the end to a continuous, layered machine process with humans routed only to high-risk decisions. From June to September, merged pull requests more than doubled while the revert rate fell by more than 40 percent. In Anthropic's monorepo, daily changes went from about 2,500 in May to more than 30,000 in September. The system runs seven layers. Before a PR even exists, Claude verifies its own work against runnable outputs — tests, builds, screenshots — then a separate subagent reviews the diff with fresh context. Bug-hunting, simplification, and end-to-end behavioral checks run as internal skills. Boris Cherny, Claude Code's creator, estimates this verification step makes results two to three times better. Two instructional files — CLAUDE.md for project conventions and REVIEW.md for area-specific failure modes — give reviewers structured context without sprawl. Once a PR opens, automated review doubled the share of PRs receiving substantive comments from 16 percent to 54 percent. Multiple agents hunt for bugs simultaneously, then a separate group votes on each finding; only consensus findings get posted, tagged as Important, Nit, or Pre-existing. Security gets its own blocking review. Standard CI — lint, type checks, platform tests, builds — runs in parallel. Every finding must be resolved before merge, either by fix or explicit dismissal with reasoning. Human review scales by risk tier, not by author identity. Low-risk changes (docs, copy, additive features behind flags) need zero human approvals once automated checks pass. Medium-risk changes need one person. The highest-risk code needs two. A classifier reads the change and assigns the tier; a second classifier can escalate. Agent approvals never count as human approvals. Whoever prompted the agent-written change is on the hook as a reviewer — the same accountability as if they typed the code themselves. The merge queue re-runs all checks against main with the change applied, catching conflicts between independently passing PRs. New behavior ships behind flags to Anthropic employees first, then to a latest channel, then stable. Feature flags can be turned off without a deploy. A weekly on-call rotation and incident reviews close the loop, asking which layer should have caught whatever slipped through. What makes this interesting is not that Anthropic uses AI to review AI code — that's table stakes. It's the explicit measurement regime: human approvals per PR, human comments, PR wait time on a person, revert rate, and incident rate are all tracked as a system. The stated goal is that most PRs merge with at most one human in the loop, with no increase in breakage. The revert-rate guardrail is the key structural choice: if automation just moves the cost from review time to incident response time, the metric catches it. The second job of code review — building shared context across a team — gets acknowledged but largely deferred. That's an honest omission, not a hidden one. The question it leaves open is whether a team that routes most reading to machines can maintain the kind of collective codebase understanding that prevents architectural drift over years, not weeks.
